Vector search is strongest when a query and passage express the same meaning with different words. Keyword search is strong for exact names, codes and rare terms. Hybrid search runs both and fuses rankings, making it a strong default candidate for mixed enterprise queries—but the choice should be proven on your evaluation set.
Where vector search helps
Dense embeddings map queries and passages into a semantic space. “Ending a subscription” can retrieve a passage titled “account cancellation.” This is valuable when users do not know the corpus vocabulary.
Semantic similarity is not factual relevance. Nearby passages can share a topic while missing the exact entity, version or constraint required.
Where exact search protects
Lexical ranking responds directly to terms. Product codes, ticket IDs, people, error strings and legal clauses often depend on exact tokens. An embedding can blur two similar identifiers that keyword retrieval keeps distinct.
Research on sparse and dense retrieval finds complementary strengths. That supports testing a hybrid—not claiming one blend wins every corpus.
How fusion works
Azure AI Search documents a common approach: run text and vector queries in parallel, then combine ranked lists with reciprocal-rank fusion. RRF rewards items that rank strongly in either list without forcing unlike raw scores onto one scale.
Fusion is one stage. Filters, reranking, freshness and source authority can also affect final evidence.
A fair evaluation
- Label expected evidence for representative questions.
- Compare keyword, vector and hybrid at the same result limit.
- Slice identifiers, paraphrases, broad topics and negatives.
- Measure recall, precision, rank and latency.
- Tune on development questions and report held-out results.
Use the RAG scorecard to record the comparison.
Built for the next step
One workspace, both kinds of signal
SPYN3 combines semantic and keyword retrieval behind a stable agent-facing contract.
Explore workspace searchSPYN3 is AI-ready data infrastructure for technical solopreneurs and AI builders. It connects business data into a workspace-isolated knowledge layer and exposes it through a unified REST API and hosted MCP server, with permission-scoped access and source-aware returns. SPYN3 is currently in private beta; current capabilities and product direction are documented on the product facts page.
Sources
- Microsoft Learn: hybrid search overview — parallel text and vector retrieval with rank fusion. Verified 13 Sep 2026
- Microsoft Learn: reciprocal rank fusion — how independent result lists are fused. Verified 13 Sep 2026
- Lin et al.: sparse, dense and hybrid retrieval — complementary strengths of sparse and dense retrieval. Verified 13 Sep 2026